Abstractive Dialogue Summarization with Sentence-Gated Modeling Optimized by Dialogue Acts

Reference

Main paper to be cited

@inproceedings{goo2018abstractive,
  title={Abstractive Dialogue Summarization with Sentence-Gated Modeling Optimized by Dialogue Acts},
    author={Chih-Wen Goo and Yun-Nung Chen},
    booktitle={Proceedings of 7th IEEE Workshop on Spoken Language Technology},
    year={2018}
}

Want to Reproduce the experiment?

Simply run python3 train.py.

Where to Put My Own Dataset?

Use --data_path=path_to_dataset.
path_to_dataset shoulde includes three folders - train, test, and valid, which is named 'train', 'test', and 'valid'.
Each of these folders contains three files - dialogue sentences, dialogue act label, and summary, which is named 'in', 'da', and 'sum'.
Each line represents an example and input sentences should be seperated by a special <EOS> token.
Vocabulary files need to be generated by yourself, including _PAD and _UNK.

Requirements

tensorflow 1.4
python 3.5

Usage

some sample usage

run with 128 units full model, and no patience for early stop
python3 train.py --num_units=128 --model_type=full --patience=0
load model and evaluate it
python3 train.py --evaluate --ckpt=full_path_to_ckpt
use "python3 train.py -h" for all avaliable parameter settings

Note

A bug found in the script of ROUGE computation is fixed, so the scores will be different from the original paper, but the trend of improvement is the same.